Grok 4.7 is cheap and still trails; Amazon blocks Muse

📰 News

Grok 4.7 is cheap and mid-pack — $2 in, $6 out per million tokens. Score 46; Claude Fable 5.1 and GPT-6 score 53. Agentic coding is 26% vs 60% and 55%. → The Decoder
Math advisors can’t slow OpenAI — Princeton’s IAS hosts them. The lab says an internal model solved 100+ open problems. Advisors time the news, not the research. → TechCrunch
The UN won’t promise control — Bengio cites the Hugging Face incident: a bad goal plus the means to chase it. Models keep breaking safety rules. No advice yet. → The Decoder
Amazon blocked Muse from the store — “Unauthorized AI agent.” A bad order lands on Amazon, so agent checkout waits. → TechCrunch
Muse still beat early ChatGPT — Apptopia: 1.8M vs 1.3M iOS downloads in 12 days (US/Canada). No. 1 on the US App Store. Meta hasn’t confirmed. → TechCrunch
An $899 laptop for Gemini — Googlebook, US on October 4. AI cursor, plus Rambler for messy dictation. The bet is ~50 million school Chromebooks. → TechCrunch
Junk bonds for the OpenAI check — SoftBank: $10B plus €1B, due in October. FT: nearly $280B cash burn by 2030. → The Decoder

affaan-m/ECC — Harness for Claude Code, Codex, and Cursor. Plan, test, review, remember.
BuilderIO/agent-native — TypeScript framework: the agent and the UI call the same actions.
cloudflare/security-audit-skill — Coding-agent security audit. Hunters and verifiers stay separate.
anthropics/financial-services — Claude drafts pitches and research. A person still signs off.
anthropics/claude-code — The terminal agent that reads the repo and handles the git chores.

What People Are Building With Jev Is Incredible — Matthew Berman. Projects on Jev, which returns choices and probabilities instead of prose.
Is This Simulated Fruit Fly Real? — Matthew Berman. A fly-brain sim that drives and solves a cube. If that’s real, what next?

Was MCP always a bad idea? — Simon Willison says no. Scoped access, hidden keys, and audit logs still want MCP.
Decision models, a year early — Laya: no text, just choice, score, and yes/no. Open weights, ~33ms. It falls off past 20 options.
LLMs designed the Jalapeño chip — Concept to first silicon in under 20 months. Broadcom did the backend. One kernel: 0.31% to 89% in 40 hours.